I need a quick and simple method to transform a PyTorch tensor with dimensions (D, M, M) into a tensor with dimensions (D // 4, M x 2, M x 2) manually without convolutions. I previously arrived to a solution for the oposite operation.(see comments on the previous question (oposite operation) )
previous question.
Quick method to transform a PyTorch tensor with gradient preservation
# Batch, Depth ,size, size -> Batch, 4 x Depth ,size//2, size//2
x = x.unfold(2, 2, 2).unfold(3, 2, 2).reshape(batch, depth, size // 2, size // 2, 4)
x = x.permute(0, 2, 3, 1, 4).reshape(batch, size // 2, size // 2, depth * 4).permute(0, 3, 1, 2)
Desired Input:
Transform (12, 2, 2) to (3, 4, 4):
[[[ 0, 1, 4, 5, 16, 17, 20, 21, 32, 33, 36, 37],
[ 2, 3, 6, 7, 18, 19, 22, 23, 34, 35, 38, 39]],
[[ 8, 9, 12, 13, 24, 25, 28, 29, 40, 41, 44, 45],
[10, 11, 14, 15, 26, 27, 30, 31, 42, 43, 46, 47]]]
Desired Output:
[[[ 0, 1, 2, 3], [ 4, 5, 6, 7], [ 8, 9, 10, 11], [12, 13, 14, 15]],
[[16, 17, 18, 19], [20, 21, 22, 23], [24, 25, 26, 27], [28, 29, 30, 31]],
[[32, 33, 34, 35], [36, 37, 38, 39], [40, 41, 42, 43], [44, 45, 46, 47]]]
EDIT: preferable working with batch. test code:
import torch
batch = 2
input = torch.arange(3*4*4*batch).reshape(batch,3, 4, 4)
batch, depth, size, _ = input.shape
x = input.unfold(2, 2, 2).unfold(3, 2, 2).reshape(batch, depth, size // 2, size // 2, 4)
x = x.permute(0, 2, 3, 1, 4).reshape(batch, size // 2, size // 2, depth * 4).permute(0, 3, 1, 2)
**reverse operation code**
print((input==x).all())
SOLUTION:
x.permute(0, 2, 3, 1).reshape(batch, size//2, size//2, depth,
2,2).permute(0, 3, 1, 4, 2, 5).reshape(batch, depth, size, size)